Tutorials

Step-by-step guides for common REINVENT4 workflows. Each tutorial includes descriptions of the input files, configuration settings in TOML file, and expected outputs.

Running REINVENT4

All workflows follow the same pattern: prepare a TOML config file, then run:

reinvent config.toml

Optional flags:

reinvent config.toml -l run.log -s 42

Flag

Description

-l FILE

Write log output to a file instead of stderr

-s N

Set random seed for reproducibility

Tutorials

  1. Sampling — generate molecules from a prior without any training or optimization.

  2. Transfer Learning — fine-tune a prior on a focused SMILES dataset to bias generation toward a target chemical series

  3. Reinforcement Learning — optimize an agent against a multi-component scoring function; includes staged/curriculum learning

  4. Scoring Function Design — how to formulate objectives, choose transforms and weights, use built-in components, and write custom ones

  5. Scoring — evaluate an existing SMILES list against a scoring function without running RL; useful for validating your scoring setup

  6. Common Workflows — end-to-end strategies combining sampling, TL, and RL for different scenarios

  7. Monitoring and Analysis — TensorBoard metrics during TL/RL, CSV output columns, and DataWarrior visualisation

Example Config Files

Ready-to-run TOML configs for each tutorial are provided in example_cfgs/:

example_cfgs/
├── sampling/   reinvent.toml, libinvent.toml, linkinvent.toml, mol2mol.toml
├── tl/         reinvent.toml, mol2mol.toml
└── rl/         single_stage.toml, multi_stage.toml

These configs reference prior model files in prior/. Download the prior models from Zenodo (record 15641297) and place them in prior/ before running.